Context (such as our location or current goal) informs everyday decisions, both by predicting stimuli and determiningrelevant responses. How do we develop priors that are general enough to apply in various contexts yet specific enough tomaximize reward in a given context? We investigated this using the AX-CPT, a task in which a cue determines which buttonto press for a probe that appears seconds later. We manipulated the frequency of the probe given the cue across participantsand built a diffusion model to estimate how the cue informs participants’ priors for the decision. We found that participants’context-dependent priors were closer to each other and less extreme than those predicted by a model that maximizes rewardrate given the true stimulus frequencies. However, participants’ priors were optimal given their subjective frequency estimates,which showed that they averaged response probabilities across cues when the cues made sufficiently similar predictions.

One of the most compelling characteristics of controlled pro-cessing is our limitation to exercise it. Theories of control allo-cation account for such limitations by assuming a cost of con-trol that constrains how much cognitive control is allocated toa task. However, this leaves open the question of why sucha cost would exist in the first place. Here, we use neural net-work simulations to test the hypothesis that constraints on cog-nitive control may reflect an optimal solution to the stability-flexibility dilemma: allocating more control to a task results ingreater activation of its neural representation but also in greaterpersistence of this activity upon switching to a new task, yield-ing switch costs. We demonstrate that constraints on controlimpair performance of any given task but reduce performancecosts associated with task switches. Critically, we show thatoptimal control constraints are higher in environments with ahigher probability of task switches.